Quality assurance · Production

Cara ships 66-ticket architecture epic autonomously using Claude Code and RePPITS structured agent methodology

The problem

Cara had 25 scaffolds as a flat list with no composition model, making every scaffold combination manual, and the AI agent safety architecture was missing — LLMs were making clinical routing decisions that should have been deterministic.

First attempt

The pre-existing system lacked a composition model, requiring every scaffold combination to be assembled manually, and LLMs were autonomously making clinical routing decisions that should have been deterministic code paths.

Workflow diagram · grounded in source
1
Architecture and ticket planning
trigger
“The architecture was complete before a single line was written. Three external model analyses synthesized with a published whitepaper on clinical safety. 66 Linear issues with blocking dependencies, interface definitions, and acceptance …”
2
RePPITS per-ticket execution
ai_action
“Every ticket followed RePPITS: Research the codebase and external sources, Propose two approaches, Plan concrete tasks, Implement, Test, then run healthcare compliance checks (HIPAA, SOC2, HITRUST) before committing. The agent cannot ski…”
3
Parallel subagent execution
ai_action
“For independent tickets, the agent spawns 2–5 parallel subagents with precise prompts”
4
Combined diff review
validation
“after parallel work, review the COMBINED diff. This rule exists because a real SSRF (Server-Side Request Forgery, where a server-side component can be tricked into making requests to unintended internal resources like AWS metadata endpoi…”
5
Phase gates
validation
“After each of 7 phases: full test suite, type-check, push, verify CI, verify Kubernetes pods. If anything fails, stop, fix, re-verify.”
6
Security audits
validation
“After Phase 4 (11,625 lines) and Phase 7 (7,708 lines), comprehensive audits ran across the entire diff. Found 2 safety-critical bugs and 10 total issues. All fixed. None deferred.”
Reported outcome

A Claude Code agent autonomously executed 66 software tickets across 2 repositories in under 4 hours, writing 536 tests and approximately 20,000 lines of code to deliver a 5-layer composable architecture; two security audits found 10 total issues including 2 safety-critical bugs, all fixed immediately with none deferred.

Reported metrics
Tickets executed autonomously66
Tests written536
Lines of code produced∼20,000 lines
Time to execute 60+ ticketsunder 4 hours
Show all 15 reported metrics
tickets executed autonomously66
tests written536
lines of code produced∼20,000 lines
time to execute 60+ ticketsunder 4 hours
parallel subagent batches19
safety-critical bugs found and fixed2
total security issues found and fixed10
production releases at milestone221
commits at milestone702
dev deployments at milestone683
active development days24 days
scaffolds after epic53
injectable files after epic113
preparation hours before autonomous executionRoughly 20–25 hours
prior infrastructure codebase size∼150K lines
Reported stack
Claude CodeClaude Opus 4.6GPT-5.4Gemini 3.1 UltraLinearMCPVS CodeCursorKubernetesS3
Source
https://widal.substack.com/p/we-shipped-a-66-ticket-architecture
Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

A Claude Code agent autonomously executed 66 software tickets across 2 repositories in under 4 hours, writing 536 tests and approximately 20,000 lines of code to deliver a 5-layer composable architecture; two security…

What tools did this team use?

Claude Code, Claude Opus 4.6, GPT-5.4, Gemini 3.1 Ultra, Linear, MCP, VS Code, Cursor, Kubernetes, S3.

What results were reported?

Tickets executed autonomously: 66; Tests written: 536; Lines of code produced: ∼20,000 lines; Time to execute 60+ tickets: under 4 hours (source-reported, not independently verified).

What failed first in this deployment?

The pre-existing system lacked a composition model, requiring every scaffold combination to be assembled manually, and LLMs were autonomously making clinical routing decisions that should have been deterministic code…

How is this quality assurance AI workflow structured?

Architecture and ticket planning → RePPITS per-ticket execution → Parallel subagent execution → Combined diff review → Phase gates → Security audits.

Related quality assurance cases